AquaRestore-UC: A Lightweight Physics-Guided Wavelet Network for Underwater Image Enhancement with an Auxiliary Uncertainty Head

Abstract

Underwater images are degraded by wavelength-dependent attenuation, backscatter, scattering-induced blur, and colour distortion. We propose AquaRestore-UC, a lightweight network combining a physics-guided restoration branch, a fixed Haar-wavelet frequency branch, and a multi-colour-space spatial branch, fused via attention gating, with an auxiliary per-pixel uncertainty head trained using a heteroscedastic loss. The uncertainty head is trained but not post-hoc calibrated; only its raw, uncalibrated coverage statistics are reported. On a completed, single-seed (seed 42) experiment at 64×64 resolution using a fixed, custom 720/80/90 UIEB split, AquaRestore-UC improves mean PSNR from 42.00 to 45.26 dB (+3.26 dB) and mean SSIM from 0.758 to 0.877 relative to raw input, across all 90 test images. Generic and parameter-matched CNN controls trained under matched conditions show smaller improvements. An exploratory, non-pre-registered paired statistical reanalysis finds these image-level differences unlikely to be due to chance, but this analysis does not establish superiority over published state-of-the-art methods, nor does it attribute the observed gains to any individual architectural component; it treats a single trained instance of each method as the unit of comparison, not multi-seed evidence. A preliminary, restoration-only three-seed extension (seeds 123 and 2024, uncertainty head untrained) is numerically consistent with the seed-42 result. Twelve of the 90 test images score lower on PSNR than raw input. We report a diagnosed and corrected heteroscedastic-uncertainty training failure and a descriptive, uncalibrated uncertainty-coverage analysis. Higher-resolution training, published-method baselines, component ablations, a fully matched multi-seed study, and post-hoc uncertainty calibration remain future work. This manuscript makes neither a state-of-the-art performance claim nor a calibrated-uncertainty claim.